Beyond the Illusion: Next-Gen Deep Learning for Deepfake Detection

Abstract

Over the past few months, it became possible to create credible face exchanges in videos, almost completely without traces of manipulation, using free deep learning-based software tools. These are called "DeepFake" videos. Digital video manipulations have been present for several decades due to the good use of visual effects. Recent advances in deep learning have increased realism in fake content and accessibility in which this can be created. AI-synthesized media, popularly referred to as "Deep Fakes", or simply "DF". Now, it has become an easy task to produce such DF using artificially intelligent tools. But the major challenge is that which concerns the detection of these DF. For the reason that training of the algorithm to spot the DF is not simple. We have gone a step further to detect the DF using Convolutional Neural Network and Recurrent neural Network. The system will use a Convolution Neural network to extract frame-level features. These characteristics are used to train an RNN, learning how to classify whether a video has been manipulated, and detecting temporal inconsistencies between the frames that are introduced by the DF creation tools. Expected result for a large set of fakes is collected from that standard data set. In this benchmark task, we demonstrate the competitiveness of our system to show results with the use of a simple architecture.